Merge branch 'master' into libretro

Author: notaz

Makefile.libretro

@@ -332,7 +332,7 @@ else ifeq ($(platform), vita)
 	CFLAGS += -fsingle-precision-constant -mword-relocations -fno-unwind-tables
 	CFLAGS += -fno-asynchronous-unwind-tables -ftree-vectorize
 	#CFLAGS += -funroll-loops # ~280K bloat
-	#CFLAGS += -fno-optimize-sibling-calls # debug?
+	CFLAGS += -fno-optimize-sibling-calls # broken arm->thumb tailcalls?
 	CFLAGS += -I$(VITASDK)/include -Ifrontend/vita
 	CFLAGS += -DNO_DYLIB
 	CFLAGS_LAST += -O3

libpcsxcore/new_dynarec/emu_if.c

@@ -342,11 +342,10 @@ static void clear_local_cache(void)
 static noinline void ari64_execute_threaded_slow(struct psxRegisters *regs,
 	enum blockExecCaller block_caller)
 {
-	if (!ndrc_g.thread.busy) {
+	if (ndrc_g.thread.busy_addr == ~0u) {
 		memcpy(ndrc_smrv_regs, regs->GPR.r, sizeof(ndrc_smrv_regs));
 		slock_lock(ndrc_g.thread.lock);
-		ndrc_g.thread.addr = regs->pc;
-		ndrc_g.thread.busy = 1;
+		ndrc_g.thread.busy_addr = regs->pc;
 		slock_unlock(ndrc_g.thread.lock);
 		scond_signal(ndrc_g.thread.cond);
 	}
@@ -357,7 +356,7 @@ static noinline void ari64_execute_threaded_slow(struct psxRegisters *regs,
 	{
 		psxInt.ExecuteBlock(regs, block_caller);
 	}
-	while (!regs->stop && ndrc_g.thread.busy && block_caller == EXEC_CALLER_OTHER);
+	while (!regs->stop && ndrc_g.thread.busy_addr != ~0u && block_caller == EXEC_CALLER_OTHER);
 
 	psxInt.Notify(R3000ACPU_NOTIFY_BEFORE_SAVE, NULL);
 	//ari64_notify(R3000ACPU_NOTIFY_AFTER_LOAD, NULL);
@@ -372,8 +371,7 @@ static void ari64_execute_threaded_once(struct psxRegisters *regs,
 		*(void **)((char *)drc_local + LO_hash_table_ptr);
 	void *target;
 
-	if (likely(!ndrc_g.thread.busy)) {
-		ndrc_g.thread.addr = 0;
+	if (likely(ndrc_g.thread.busy_addr == ~0u)) {
 		target = ndrc_get_addr_ht_param(hash_table, regs->pc,
 				ndrc_cm_no_compile);
 		if (target) {
@@ -412,21 +410,21 @@ static void ari64_execute_threaded_block(struct psxRegisters *regs,
 
 static void ari64_thread_sync(void)
 {
-	if (!ndrc_g.thread.lock || !ndrc_g.thread.busy)
+	if (!ndrc_g.thread.lock || ndrc_g.thread.busy_addr == ~0u)
 		return;
 	for (;;) {
 		slock_lock(ndrc_g.thread.lock);
 		slock_unlock(ndrc_g.thread.lock);
-		if (!ndrc_g.thread.busy)
+		if (ndrc_g.thread.busy_addr == ~0)
 			break;
 		retro_sleep(0);
 	}
 }
 
 static int ari64_thread_check_range(unsigned int start, unsigned int end)
 {
-	u32 addr = ndrc_g.thread.addr;
-	if (!addr)
+	u32 addr = ndrc_g.thread.busy_addr;
+	if (addr == ~0u)
 		return 0;
 
 	addr &= 0x1fffffff;
@@ -451,16 +449,17 @@ static void ari64_compile_thread(void *unused)
 	slock_lock(ndrc_g.thread.lock);
 	while (!ndrc_g.thread.exit)
 	{
-		if (!ndrc_g.thread.busy)
+		addr = *(volatile unsigned int *)&ndrc_g.thread.busy_addr;
+		if (addr == ~0u)
 			scond_wait(ndrc_g.thread.cond, ndrc_g.thread.lock);
-		addr = ndrc_g.thread.addr;
-		if (!ndrc_g.thread.busy || !addr || ndrc_g.thread.exit)
+		addr = *(volatile unsigned int *)&ndrc_g.thread.busy_addr;
+		if (addr == ~0u || ndrc_g.thread.exit)
 			continue;
 
 		target = ndrc_get_addr_ht_param(hash_table, addr,
 				ndrc_cm_compile_in_thread);
 		//printf("c  %08x -> %p\n", addr, target);
-		ndrc_g.thread.busy = 0;
+		ndrc_g.thread.busy_addr = ~0u;
 	}
 	slock_unlock(ndrc_g.thread.lock);
 	(void)target;
@@ -490,7 +489,7 @@ static void ari64_thread_shutdown(void)
 		slock_free(ndrc_g.thread.lock);
 		ndrc_g.thread.lock = NULL;
 	}
-	ndrc_g.thread.busy = ndrc_g.thread.addr = 0;
+	ndrc_g.thread.busy_addr = ~0u;
 }
 
 static void ari64_thread_init(void)
@@ -514,7 +513,8 @@ static void ari64_thread_init(void)
 		return;
 
 	ari64_thread_shutdown();
-	ndrc_g.thread.busy = ndrc_g.thread.addr = ndrc_g.thread.exit = 0;
+	ndrc_g.thread.exit = 0;
+	ndrc_g.thread.busy_addr = ~0u;
 
 	if (enable) {
 		ndrc_g.thread.lock = slock_new();

libpcsxcore/new_dynarec/linkage_arm.S

@@ -165,7 +165,7 @@ FUNCTION(dyna_linker):
 	beq	0f
 	add	r6, r5, r6, asr #6  /* old target */
 	teq	r0, r6
-	moveq	pc, r0 /* Stale i-cache */
+	bxeq	r0     /* Stale i-cache */
 	mov	r0, r4
 	mov	r1, r6
 	bl	ndrc_add_jump_out
@@ -176,15 +176,15 @@ FUNCTION(dyna_linker):
 	sub	r1, r1, #2
 	add	r1, r1, r2, lsr #8
 	str	r1, [r5]
-	mov	pc, r8
+	bx	r8
 0:
 	mov	r0, r4
 #else
 	/* XXX: should be able to do better than this... */
 #endif
 	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
-	mov	pc, r0
+	bx	r0
 	.size	dyna_linker, .-dyna_linker
 
 	.align	2
@@ -234,7 +234,7 @@ FUNCTION(jump_vaddr_r7):
 FUNCTION(jump_vaddr_r0):
 	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
-	mov	pc, r0
+	bx	r0
 	.size	jump_vaddr_r0, .-jump_vaddr_r0
 
 	.align	2
@@ -257,10 +257,10 @@ FUNCTION(cc_interrupt):
 	tst	r2, r2
 	ldmfdne	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, pc}
 	cmp	r0, r9
-	moveq	pc, lr
+	bxeq	lr
 	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
-	mov	pc, r0
+	bx	r0
 	.size	cc_interrupt, .-cc_interrupt
 
 	.align	2
@@ -317,7 +317,7 @@ FUNCTION(jump_to_new_pc):
 	bne	new_dyna_leave
 	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
-	mov	pc, r0
+	bx	r0
 	.size	jump_to_new_pc, .-jump_to_new_pc
 
 	.align	2
@@ -427,7 +427,7 @@ new_dyna_start_at_e:
 	ldr	r10, [fp, #LO_cycle]
 	str	r1, [fp, #LO_last_count]
 	sub	r10, r10, r1
-	mov	pc, r0
+	bx	r0
 	.size	new_dyna_start, .-new_dyna_start
 
 /* --------------------------------------- */
@@ -456,7 +456,7 @@ new_dyna_start_at_e:
 .else
 	\readop	r0, [r1, r3, lsl #\tab_shift]
 .endif
-	movcc	pc, lr
+	bxcc	lr
 	mov	r2, r12
 	str	r12, [fp, #LO_cycle]
 .endm
@@ -495,7 +495,7 @@ FUNCTION(jump_handler_read32):
 .else
 	\wrtop	r1, [r3, r12, lsl #\tab_shift]
 .endif
-	movcc	pc, lr
+	bxcc	lr
 	ldr	r12, [fp, #LO_last_count]
 	mov     r0, r1
 	add	r2, r2, r12

libpcsxcore/new_dynarec/new_dynarec.c

@@ -9010,7 +9010,7 @@ static struct block_info *new_block_info(u_int start, u_int len,
   return block;
 }
 
-static int new_recompile_block(u_int addr)
+static int noinline new_recompile_block(u_int addr)
 {
   u_int pagelimit = 0;
   u_int state_rflags = 0;

libpcsxcore/new_dynarec/new_dynarec.h

@@ -24,8 +24,7 @@ struct ndrc_globals
 		void *cond;
 		void *dirty_start;
 		void *dirty_end;
-		unsigned int addr;
-		int busy;
+		unsigned int busy_addr; // 0 is valid, ~0 == none
 		int exit;
 	} thread;
 };

libpcsxcore/psxinterpreter.c

@@ -1243,7 +1243,8 @@ static void intNotify(enum R3000Anote note, void *data) {
 		setupCop(psxRegs.CP0.n.SR);
 		// fallthrough
 	case R3000ACPU_NOTIFY_CACHE_ISOLATED: // Armored Core?
-		memset(&ICache, 0xff, sizeof(ICache));
+		if (fetch == fetchICache)
+			memset(&ICache, 0xff, sizeof(ICache));
 		break;
 	case R3000ACPU_NOTIFY_CACHE_UNISOLATED:
 		break;
@@ -1340,8 +1341,10 @@ void intApplyConfig() {
 
 	// the dynarec may occasionally call the interpreter, in such a case the
 	// cache won't work (cache only works right if all fetches go through it)
-	if (!Config.icache_emulation || psxCpu != &psxInt)
+	if (!Config.icache_emulation || psxCpu != &psxInt) {
 		fetch = fetchNoCache;
+		memset(&ICache, 0xff, sizeof(ICache));
+	}
 	else
 		fetch = fetchICache;